Yeah, the bug is unrelated to this patch. I'll open a PR against netmap and won't let it block committing work! And thank you very much for being willing to commit this patch! ^_^

I'm testing on the master branch of the repo cloned from https://github.com/freebsd/freebsd.git with GENERIC conf on a desktop computer. And I can also reproduce this panic in VirtualBox with the image downloaded from http://mirrors.aliyun.com/freebsd/snapshots/VM-IMAGES/11.0-CURRENT/amd64/20151217/FreeBSD-11.0-CURRENT-amd64-20151217-r292413.vhd.xz

I tried to test this patch on -HEAD, but it caused kernel panic. It seems that there is a bug in the kqueue support code of netmap. And I can reproduce the kernel panic with this simple program on -HEAD:

In D3563#79576, @adrian wrote:

ok, this looks good. I'll see if I can get gnn / bz to review this before I commit it to -HEAD.

Thanks very much for digging into this!

Replace the error handling code in ip6_direct_input() with KASSERT(), because this tag is mandatory.

I mean this line!!!

Oops! I didn't mean this line!

In D3563#79549, @adrian wrote:

Right - so it may have a new src/dst L3/L4 which means it needs to go back through the stack.

So as long as it's going back through and getting re-hashed on its trip though, I think we're okay.

How about adding a counter for when we drop frames on that direct dispatch v6 netisr so people can see when things are failing? Then I think we'll be ready.

I have defined a struct to save the context info for reinjecting. Many thanks! ^_^

Define a struct to save the context info for reinjecting.

Use an mbuf tag instead of the mis-understood fields.

Fix various bugs.

Remove some obsolete comments about the flowid of directly dispatched frames.
Now the flowid of directly dispatched frames will be recalculated by
rss_soft_m2cpuid_v4() if needed. So there is no such assumption any more.

This patch requires more work. I'm so sorry!!! >_<

Fix the bug related to atomic fragments in the previous patch
Atomic fragments should also be dispatched to ip6_direct.

Use ppsratelimit() for doing message rate limiting and and a global rate limit (maxpps, ie. the 3rd parameter of ppsratelimit()) for RSS.

In D3471#71596, @adrian wrote:

Actually, we've hand-rolled 'ratecheck()'. Would you like to modify your code to use ppsratecheck() instead? That'd make this all much easier.

In D3423#70070, @gnn wrote:

Shouldn't the printf() be under a debug statement?